
%matplotlib inline

import matplotlib.pyplot as plt
import pandas as pd
import numpy as np
from PIL import Image
import matplotlib.image as mpimg
from tqdm import tqdm

from IPython.display import HTML, display


train_features = pd.read_csv("../competition_data/train_features.csv", index_col="id")
test_features = pd.read_csv("../competition_data/test_features.csv", index_col="id")
train_labels = pd.read_csv("../competition_data/train_labels.csv", index_col="id")

display(train_features.head())
display(train_labels.head())


# reverse One-Hot-Encoding
species_labels = sorted(train_labels.columns.unique())

train_labels_cat = train_labels.copy()
train_labels_cat['label'] = train_labels[species_labels].idxmax(axis=1)
train_labels_cat = train_labels_cat.drop(species_labels, axis=1)

display(train_labels_cat.head())


print(species_labels)
print(f'Anzahl Tierklassen: {len(species_labels)}')

['antelope_duiker', 'bird', 'blank', 'civet_genet', 'hog', 'leopard', 'monkey_prosimian', 'rodent']
Anzahl Tierklassen: 8


img = mpimg.imread('class_images/class_images.jpg')
plt.imshow(img)
plt.axis('off')
plt.show()


random_state = 42
path_img = '../competition_data/'

# we'll create a grid with 8 positions, one for each label (7 species, plus blanks)
fig, axes = plt.subplots(nrows=4, ncols=2, figsize=(20, 20))

# iterate through each species
for species, ax in zip(species_labels, axes.flat):
    # get an image ID for this species
    img_id = (
        train_labels[train_labels.loc[:,species] == 1]
        .sample(1, random_state=random_state)
        .index[0]
    )
    # reads the filepath and returns a numpy array
    img = mpimg.imread(path_img + train_features.loc[img_id].filepath)
    # plot etc
    ax.imshow(img)
    ax.set_title(f"{img_id} | {species}")


# we'll create a grid with 8 positions, one for each label (7 species, plus blanks)
fig, axes = plt.subplots(nrows=4, ncols=2, figsize=(20, 20))

# iterate through each species
for species, ax in zip(species_labels, axes.flat):
    # get an image ID for this species
    img_id = (
        train_labels[train_labels.loc[:,species] == 1]
        .sample(1)
        .index[0]
    )
    # reads the filepath and returns a numpy array
    img = mpimg.imread(path_img + train_features.loc[img_id].filepath)
    # plot etc
    ax.imshow(img)
    ax.set_title(f"{img_id} | {species}")


example_bad_img = ['ZJ015580', 'ZJ002746', 'ZJ007054', 'ZJ000888', 'ZJ010341', 'ZJ014451',
                   'ZJ004927', 'ZJ007091', 'ZJ013234', 'ZJ013093', 'ZJ004451', 'ZJ010190',
                   'ZJ002138', 'ZJ002196']
example_good_img = ['ZJ003890', 'ZJ004925', 'ZJ012762', 'ZJ014396', 'ZJ015264', 'ZJ000895',
                    'ZJ007334', 'ZJ004978', 'ZJ014157', 'ZJ010885']


def plot_image_from_image_id(image_ids: list, nrows=4, ncols=3, figsize=(15, 15),
                             fontsize=10, path='../competition_data/'):
    # create grid 
    fig, axes = plt.subplots(nrows=nrows, ncols=ncols, figsize=figsize)

    # iterate through each bad image
    for idx, (img_id, ax) in enumerate(zip(image_ids, axes.flat)):
        # get image label
        img_label = train_labels_cat.loc[image_ids[idx]]['label']
        # reads the filepath and returns a numpy array
        img = mpimg.imread(path + train_features.loc[img_id].filepath)
        # image dimension
        width, height = Image.open(path_img + train_features.loc[img_id].filepath).size
        # plot etc
        ax.imshow(img)
        ax.set_title(f"{img_id} | {img_label} | {str(width)}x{str(height)}", fontsize=fontsize)

plot_image_from_image_id(example_bad_img)
#plot_image_from_image_id(example_good_img)


display(train_labels.sum().sort_values(ascending=False))
display(train_labels.sum().divide(train_labels.shape[0]).sort_values(ascending=False))

monkey_prosimian    2492.0
antelope_duiker     2474.0
civet_genet         2423.0
leopard             2254.0
blank               2213.0
rodent              2013.0
bird                1641.0
hog                  978.0
dtype: float64

monkey_prosimian    0.151140
antelope_duiker     0.150049
civet_genet         0.146955
leopard             0.136705
blank               0.134219
rodent              0.122089
bird                0.099527
hog                 0.059316
dtype: float64


train_features.head(2)


print(f'Anzahl Standort Training-Kameras: {len(train_features.site.unique())}')  
print(f'Anzahl Standort Test-Kameras: {len(test_features.site.unique())}')

print(f'Mittelwert: {len(train_features.site) / len(train_features.site.unique()):.0f} Bilder pro Ort, Train')
print(f'Mittelwert: {len(test_features.site) / len(test_features.site.unique()):.0f} Bilder pro Ort, Test')

Anzahl Standort Training-Kameras: 148
Anzahl Standort Test-Kameras: 51
Mittelwert: 111 Bilder pro Ort, Train
Mittelwert: 88 Bilder pro Ort, Test


df_site_count = train_features.groupby('site').count().reset_index().sort_values('filepath', ascending=True)

# plot histgram
plt.figure(figsize=(20, 10))
plt.bar(df_site_count.site, df_site_count.filepath)
plt.title('Anzahl Bilder je Kamera Standort (Train)')
plt.xlabel('site ID')
plt.ylabel('counts')
plt.xticks(rotation=90)
plt.grid()
plt.show()


df_train_feature_labels = train_features.merge(train_labels_cat, left_index=True, right_index=True)
df_train_feature_labels.head()


df_stacked_plot = pd.crosstab(df_train_feature_labels['site'], df_train_feature_labels['label'])
df_stacked_plot['max_count'] = df_stacked_plot.sum(axis=1)
df_stacked_plot = df_stacked_plot.sort_values('max_count', ascending=True)
df_stacked_plot.head()


def plot_stacked_classes_site(stacked_plot, min_count = 50):
    stacked_plot = stacked_plot[stacked_plot['max_count'] > min_count]
    stacked_plot = stacked_plot.drop(columns='max_count')


    stacked_plot.plot(kind='bar', stacked=True, figsize=(20,10))
    plt.title('Vorkommen Tierklassen je Standort (Train)')
    plt.xlabel('site ID')
    plt.ylabel('counts')
    plt.xticks(rotation=90)
    plt.grid()
    plt.show()

plot_stacked_classes_site(df_stacked_plot, 50)


train_features[train_features['site'] == 'S0007']


def plot_image_from_site_id(site_id: str, nrows=4, ncols=3, figsize=(15, 15), path='../competition_data/'):
    # get images from site
    img_site = train_features[train_features['site'] == site_id].reset_index()
    img_site = img_site[0:nrows*ncols]
    # create grid 
    fig, axes = plt.subplots(nrows=nrows, ncols=ncols, figsize=figsize)

    # iterate through each bad image
    for img_id, image_path, ax in zip(img_site.id, img_site.filepath, axes.flat):
        # reads the filepath and returns a numpy array
        img = mpimg.imread(path + image_path)
         # get image label
        img_label = train_labels_cat.loc[img_id]['label']
        # plot
        ax.imshow(img)
        ax.set_title(f"{img_id} | {img_label}")

plot_image_from_site_id('S0007')


mode_to_bpp = {'1':1, 'L':8, 'P':8, 'RGB':24, 'RGBA':32, 'CMYK':32, 'YCbCr':24, 'I':32, 'F':32}

img_id = []
img_width = []
img_height = []
img_dim = []
band_mode = []
bit_depth = []

for imag_id in train_features.reset_index().id:
    # get id and dimensions
    img = Image.open(path_img + train_features.loc[imag_id].filepath)
    width, height = img.size
    dim = img.size
    mode = img.getbands()

    # save to list
    img_id.append(imag_id)
    img_width.append(width)
    img_height.append(height)
    img_dim.append(dim)
    band_mode.append(mode)

df_img_shape_pil = pd.DataFrame({'id': img_id, 'width': img_width, 'height': img_height, 
                                 'dim':img_dim, 'band_mode': band_mode})

df_img_shape_pil.head(2)


unique_dim = df_img_shape_pil.dim.value_counts().reset_index().sort_values('dim', ascending=False)
unique_dim = unique_dim.rename(columns={'dim': 'count', 'index': 'dim'})
unique_dim['count_relativ'] = np.round(unique_dim['count'] / sum(unique_dim['count']), 2)
#display(unique_dim)


fig, ax = plt.subplots(1, 2, figsize=(8, 5))

ax[0].bar(np.arange(len(unique_dim)), unique_dim['count'])
ax[0].set_xticks(np.arange(len(unique_dim)), unique_dim.dim)
ax[0].set_title('Verteilung der Bildauflösungen')
ax[0].set_ylabel('counts')
ax[0].set_xticklabels(unique_dim.dim, rotation=60)

for i, v in enumerate(unique_dim['count']):
    ax[0].text(i, v, str(v), ha='center', va='bottom')

ax[1].bar(np.arange(len(unique_dim)), unique_dim['count_relativ'] )
ax[1].set_xticks(np.arange(len(unique_dim)), unique_dim.dim)
ax[1].set_title('Verteilung der Bildauflösungen [relativ]')
ax[1].set_ylabel('counts')
ax[1].set_xticklabels(unique_dim.dim, rotation=60)


for i, v in enumerate(unique_dim['count_relativ']):
    ax[1].text(i, v, str(v), ha='center', va='bottom')

plt.show()


df_bit_depth = df_img_shape_pil['band_mode'].value_counts().reset_index(name='count')
df_bit_depth = df_bit_depth.rename(columns={'index':'mode'})
df_bit_depth.plot(x='mode', y='count', kind='bar')
plt.suptitle('Verteilung Bit Tiefe Bilder', fontsize=12)
plt.title('RGB: 24bit, L: 8bit', fontsize=8)
plt.xlabel('band mode')
plt.ylabel('count')
plt.xticks(rotation=0)
plt.grid()
plt.show()


# images with Logo and time:
img_logo_time = ['ZJ000001', 'ZJ000002', 'ZJ000003', 'ZJ000005',
                 'ZJ000006', 'ZJ000025', 'ZJ000026', 'ZJ000031', 'ZJ000140']


plot_image_from_image_id(img_logo_time, nrows=3, figsize=(10, 7), fontsize=8)


# images cut
img_logo_time_cut = ['ZJ000004', 'ZJ000053', 'ZJ000063', 'ZJ000090',
                 'ZJ000099', 'ZJ000106', 'ZJ000114', 'ZJ000119', 'ZJ000156']

plot_image_from_image_id(img_logo_time_cut, nrows=3, figsize=(10, 7), fontsize=8)


# special cases
img_special_case = ['ZJ000019', 'ZJ000015', 'ZJ000016', 'ZJ000018',
                 'ZJ000065', 'ZJ000124', 'ZJ000132', 'ZJ000139', 'ZJ000142']

plot_image_from_image_id(img_special_case, nrows=3, figsize=(10, 7), fontsize=8)


df_site_count[df_site_count['filepath'] > 400]


plot_image_from_site_id('S0106')


# crop image
def crop_image_for_logo(pillow_image, plot_image=False):
    # dim img
    width, height = img.size

    # crop images, left corner
    crop_height = height * 0.09
    crop_width = width * 0.95

    x1 = 0
    y1 = height - crop_height
    x2 = width - crop_width
    y2 = height

    cropped_img = img.crop((x1, y1, x2, y2))

    if plot_image:
        plt.imshow(cropped_img)
        plt.show()

    return cropped_img

# get image main color
def get_img_main_color(pillow_imag, print_info=False):
    # Wandle das Bild in den RGB-Modus um, falls es noch nicht im RGB-Modus vorliegt.
    if pillow_imag.mode != "RGB":
        pillow_imag = pillow_imag.convert("RGB")

    # Extrahiere die Farb-Kanäle aus dem Bild.
    r, g, b = pillow_imag.split()[0], pillow_imag.split()[1], pillow_imag.split()[2]

    # Berechne den durchschnittlichen Wert des Kanals.
    mean_r = sum(r.getdata()) / len(r.getdata())
    mean_g = sum(g.getdata()) / len(g.getdata())
    mean_b = sum(b.getdata()) / len(b.getdata())

    mean_rgb = {'red': mean_r, 'green': mean_g, 'blue': mean_b}

    max_mean_color = max(mean_rgb, key=lambda x:mean_rgb[x])
    if print_info:
        print(f'image is mostly: {max_mean_color}')
        return
    
    return max_mean_color


# read image
img = Image.open(path_img + train_features.loc['ZJ000001'].filepath)

# crop image
cropped_image = crop_image_for_logo(img, plot_image=True)

# get main color image
get_img_main_color(cropped_image, print_info=True)

image is mostly: red

	filepath	site
id
ZJ000000	train_features/ZJ000000.jpg	S0120
ZJ000001	train_features/ZJ000001.jpg	S0069
ZJ000002	train_features/ZJ000002.jpg	S0009
ZJ000003	train_features/ZJ000003.jpg	S0008
ZJ000004	train_features/ZJ000004.jpg	S0036

	filepath	site
id
ZJ000000	train_features/ZJ000000.jpg	S0120
ZJ000001	train_features/ZJ000001.jpg	S0069

	filepath	site
id
ZJ003948	train_features/ZJ003948.jpg	S0007
ZJ006104	train_features/ZJ006104.jpg	S0007
ZJ007333	train_features/ZJ007333.jpg	S0007
ZJ007590	train_features/ZJ007590.jpg	S0007
ZJ008782	train_features/ZJ008782.jpg	S0007
ZJ010893	train_features/ZJ010893.jpg	S0007
ZJ012885	train_features/ZJ012885.jpg	S0007
ZJ013315	train_features/ZJ013315.jpg	S0007
ZJ013350	train_features/ZJ013350.jpg	S0007

	id	width	height	dim	band_mode
0	ZJ000000	960	540	(960, 540)	(R, G, B)
1	ZJ000001	960	540	(960, 540)	(R, G, B)

	site	filepath
88	S0120	423
32	S0038	429
45	S0059	438
34	S0043	444
31	S0036	456
7	S0008	541
49	S0063	557
8	S0009	664
46	S0060	1132

Deep Learning: Was versteckt sich da?¶

Explorative Datenanalyse¶

CSV Daten Lesen (Train, Test, Label)¶

Die Klassifikation Tierarten¶

Bilddaten Lesen (Train)¶

Random Ansicht¶

Problematische Bilder¶

Verteilungen der Tierklassen¶

Analyse zu Kamera Standorte (site ID)¶

Verteilung der Bilder je Standort (Train)¶

Verteilung der Tierklassen je Standort¶

Plotten der Bilder je Standort¶

Analyse zu Bildauflösung¶

Bilder und Bittiefe¶

Spezifische Kameramerkmale¶

Bild bearbeitung durch zuschneiden¶

Weitere Bildauflösungs Merkmale¶

Analysieren von Kamerastandorte¶

Augmentation Test¶

Tests¶

	label
id
ZJ000000	bird
ZJ000001	monkey_prosimian
ZJ000002	bird
ZJ000003	monkey_prosimian
ZJ000004	leopard

label	antelope_duiker	bird	blank	civet_genet	hog	leopard	monkey_prosimian	rodent	max_count
site
S0102	0	0	0	0	0	0	1	0	1
S0078	0	0	2	0	0	0	0	0	2
S0079	0	0	2	0	0	0	0	0	2
S0178	0	0	0	0	0	2	0	0	2
S0143	1	0	2	0	0	0	0	0	3

	antelope_duiker	bird	blank	civet_genet	hog	leopard	monkey_prosimian	rodent
id
ZJ000000	0.0	1.0	0.0	0.0	0.0	0.0	0.0	0.0
ZJ000001	0.0	0.0	0.0	0.0	0.0	0.0	1.0	0.0
ZJ000002	0.0	1.0	0.0	0.0	0.0	0.0	0.0	0.0
ZJ000003	0.0	0.0	0.0	0.0	0.0	0.0	1.0	0.0
ZJ000004	0.0	0.0	0.0	0.0	0.0	1.0	0.0	0.0